15 copy-and-paste prompts to test whether ChatGPT can find your products

Ask ChatGPT for a product recommendation in your category and watch what comes back. It finds some products, names some merchants, and gets a detail wrong. The price is last month's. It says to check store availability rather than answering. Or your product is not in the answer at all.

None of that is a ranking problem. It is a retrieval problem, and unlike ranking it is measurable. The facts about your products either can or cannot be reached by the system assembling the answer, and the point where that breaks is specific enough to find in an afternoon.

This is the field test we run with clients. Fifteen prompts, a scorecard for what came back, and a map from each failure to the layer most likely responsible.

What's in the AI product catalogue retrieval test

On this page you get the five stages of retrieval, the five categories the test covers, and one example prompt from each. The full PDF has all fifteen and the diagnostic.

• 15 copy-and-paste prompts across discovery, price and inventory, relationships, product knowledge and a deliberately difficult stress test
• Five versions of that stress test, one each for multi-store retail, B2B and wholesale, parts and compatibility, multi-region, and programs and services
• A scorecard covering discovery, attribute accuracy, variant accuracy, price, availability, fulfilment, relationships and source quality
• A map from each failed prompt to the retrieval stage most likely responsible
• Guidance on re-testing, because the before-and-after comparison is worth more than a single result

Get the PDF using the form on this page.

How do you rank in ChatGPT?

You do not rank in ChatGPT the way you rank in Google. There is no position to climb and no results page to move up. When someone asks for a product, the system assembles an answer from whatever it can find, understand and trust at that moment.

Three things decide whether you appear in it. It has to find the product in answer to a question that never mentions your brand. It has to understand the facts, which means typed attributes rather than adjectives buried in prose. And it has to trust the source, which means the price it quotes is the price you are charging today and the stock it promises is stock you actually have.

A page that looks complete to a person can still be invisible to retrieval. A person can read a specification table, click through to store availability, and combine three things on screen. A retrieval system reads what is exposed to it and nothing else.

Why can't AI find your products?

When AI misses something you sell, the failure almost always sits at one point along the path from a page a person can read to a source a machine can query. The stages run in order, and finding where yours stops is the whole diagnosis.

• Human-visible. A customer can find the information by using the interface.
• Crawlable. A retrieval system can reliably fetch it.
• Structured. It exists as explicit commercial data rather than being inferred from prose.
• Channel-ingested. It is supplied through a feed or data source the destination supports.
• Integrated for retrieval. A live service or API is explicitly connected to the system answering the query.

Most teams are at the first stage and assume they are at the third. Store-level stock behind a find-in-store button, compatibility held in a PDF, and a specification table rendered by the CMS but never stored as fields are the three most common examples.

So the useful question is not why an AI did not rank you. It is whether the business can answer the question internally, and if so, where the retrieval path stops. Two or more failures at the same stage is a platform problem rather than a content problem.

How do you test whether AI can find your products?

Pick 5 to 10 representative products and include at least one difficult one, with multiple variants, local stock, accessories, compatibility rules or complex fulfilment. Use a clean or logged-out session. Then compare each answer against the commercial truth in your own systems.

The test covers five categories. One example from each is below, and the rest are in the PDF.

Discovery

Whether an AI can find your product at all when the customer never types your brand name.

I need a [product category] under [budget] for [specific use case]. What would you recommend and where can I buy it?

Check whether your product appears without the brand being named, which competitors appear instead, and which sources get cited.

Price, inventory and fulfilment

The facts that change. These are the answers most likely to be confidently wrong, because they were true when they were last published.

I'm in [suburb or postcode]. Where can I buy [product] in stock today?

Check whether it resolves store-level availability, or stops at a generic instruction to check store availability.

Relationships and compatibility

What works with what. Conversational shopping produces relational questions constantly, and most catalogues hold the answer in a PDF or a diagram.

I already own [model or product]. Which [part or accessory] is compatible with it?

Check whether the relationship is correct and supported by an authoritative source.

Product knowledge

The specifics a buyer asks before committing. This is where an answer sounds authoritative and is subtly wrong.

Does [product] support [technical requirement]? Give me the exact specification and source.

Check whether the answer is grounded in an explicit specification, manual or structured field.

The stress test

The most useful test in the guide. It forces identity, attributes, relationships, price, inventory, location and fulfilment to resolve in one answer, and the point where it breaks tells you which layer to fix.

I'm in [location]. I need [product category] under [budget], with [attribute], compatible with [existing product], available for pickup by [date]. What are my best options?

The PDF has five versions of this one, because the difficult query is not the same for a distributor, a franchise network and a membership organisation.

This is not a ranking test. A product missing from one answer does not prove your catalogue is broken. Treat it as a blind-spot diagnostic for discovery, structure, price, inventory, fulfilment and relationships, and run the same prompts again after you change something.

Who this test is for

• Ecommerce and digital leads who need to know whether AI channels are reaching their catalogue before peak season
• Merchandising and product data owners responsible for attributes, variants and feed quality
• SEO and content teams extending their remit from pages to product data
• Platform and architecture teams deciding where commercial truth should live

Your catalogue is becoming an interface

The goal is not to build a separate catalogue for ChatGPT, then another for Google, then another for whatever launches next. That recreates the problem every time an interface appears, and guarantees the versions drift apart.

Model the commercial truth once, covering identity, attributes, variants, price, availability, relationships and fulfilment, then make each retrieval channel an output of it.

The full research behind this test is in product feed optimization for AI shopping, which covers what OpenAI and Google currently ask merchants to structure. For the paid side of the same surface, see ChatGPT ads and eCommerce.